审核图片和视频 — 发布前需检查的事项

简短回答。 首先，请问帖子是否真的需要这张图片。如果需要，请检查截图中的每张脸、每个名字和号码，任何文件或标签，以及任何显示个人居住或工作地点的信息。观看视频时请务必看完并打开声音。自动筛选功能会检查露骨和暴力内容；它不会检查隐私问题。

A text submission says what its author chose to say. An image says a good deal that its author never noticed. The reflection in a window, the name at the top of a chat, the lanyard round someone's neck — a picture identifies people and places with a thoroughness no paragraph manages, and an admin who looks only at what the image is of will miss most of it.

Automated screening helps with one part of this problem. It is fairly good at telling you whether an image is sexually explicit or violent. It is not looking for privacy problems at all: a perfectly sharp photograph of a stranger's identity card is, to a content-safety model, a completely safe image. That part of the check belongs to you.

Start by asking whether the image is needed

Many submissions attach an image out of habit. If the text stands on its own, publish the text. Every image you do not publish is a set of risks you did not have to assess. A sensible default for a confession page is that images are the exception and must earn their place.

People in the frame

Who is identifiable? Look at every face, not only the subject's. Look for name tags, lanyards, jerseys with names, uniforms and school badges.

Did they agree to this? The person who submitted the photo may be in it. The other people did not submit it. A picture of someone asleep in a lecture, eating alone, or caught mid-fall is funny to everyone except them.

Could it embarrass or endanger them? Photos taken without the subject's knowledge — in class, on public transport, in a gym — should be declined almost without exception. Any photograph taken in a toilet, changing room, bedroom or hostel room without clear consent is declined always.

Is anyone under 18? Images of children and teenagers need extra care. If you cannot be confident that the young person and a parent would be content to see it published, do not publish it.

If the image is worth publishing and the only problem is a bystander, crop them out or blur them. Blur heavily — a light blur over a face is still recognisable to anyone who knows the person — and look at the result at full size before posting.

Screenshots of conversations

Chat screenshots are the commonest image in a confession queue and the most consistently mishandled. Check all of the following:

• The header — the contact's name, their photo, the group's name.
• Profile pictures beside each message.
• Phone numbers, which appear in place of names for people who are not saved contacts.
• Names used inside the messages, including nicknames.
• The status bar — the clock, battery level and network name can show whose phone it was and when.
• Notification banners that happened to drop down as the screenshot was taken.
• Other chats visible in a sidebar or behind a pop-up.

Remember, too, that a private conversation has at least two participants, and only one of them sent it to you. Publishing someone's private messages to shame them is a harm in its own right, even with the name scribbled out — the other person will recognise their own words, and so will their friends.

If a conversation matters to the story, the safer course is to ask for it retold in the submitter's own words.

Documents and screens

Decline, or crop thoroughly, anything that shows:

• identity cards, passports, student or staff cards;
• results slips, transcripts or offer letters;
• payslips, bank statements, receipts showing a name or card number;
• medical letters, appointment cards, prescriptions;
• delivery labels with a name, address or telephone number;
• tickets and boarding passes, including their barcodes and QR codes;
• a laptop or phone screen in the background of some other photograph.

A scribble made with a phone's markup pen is often not enough. Semi-transparent highlighters let the text show through, and thin strokes leave the outlines of the digits legible. Use a solid opaque block, or crop the area out completely.

Where the picture was taken

A photograph can reveal a location in two ways.

What is visible. A street sign, a shop front, a distinctive building, a view from a particular window, a door number. If the post concerns a person and the image shows where they live or work, you may be about to publish their address.

What is hidden in the file. Photos from phones can carry embedded metadata — the date, the device, and sometimes the exact coordinates where the photo was taken. Large platforms generally remove this when an image is uploaded, but if you host images yourself, or pass the original file around your team, the data travels with it. A screenshot of the photo, rather than the original, carries none of it.

Vehicle registration plates fall into the same category. A plate together with a location and a complaint about someone's parking amounts to a public invitation to go and find that car.

Sexual and violent content

This is where automated screening earns its place, and you should let it take the first look, so that fewer people on your team have to see the worst material.

Explicit images of adults have no place on a community page whatever the context, and intimate images shared without the subject's consent are a serious harm and may be a criminal offence. Decline and delete. Do not forward them to other admins "to check".

Sexual images of anyone under 18 are illegal to possess or share in Malaysia and almost everywhere else. If one arrives:

• do not save it, take a screenshot of it, or send it to anyone, including to another admin or as "evidence";
• report it through the hosting platform's reporting tools and to the police;
• follow the instructions you are given about the file.

Violence and gore — accidents, fights, injuries. Ask what publishing achieves. Footage of a fight on campus identifies everyone in it and is rarely anything but spectacle.

Manipulated and AI-generated images

It is now trivial to produce a convincing fake photograph or to place a real person's face on another body. Assume that any image which is both surprising and damaging to someone may be fabricated.

• Sexualised fakes of real people are declined and deleted, exactly as real intimate images would be. The harm to the person depicted is the same.
• Fake screenshots of chats, posts or documents are easy to make and impossible to verify by looking at them. A screenshot is not evidence that someone said something.
• Edited images intended as satire of public figures are a matter of judgement; of private individuals, they are simply harassment.

You will not always be able to tell what is real. That uncertainty is itself a reason to decline.

视频

Everything above applies to video, with three additional problems.

You have to watch all of it. The thumbnail and the first five seconds tell you nothing about the forty-fifth. If you do not have time to watch a clip to the end, you do not have time to publish it.

Listen as well as look. Names are spoken, people are addressed, and voices are recognisable. A clip that looks anonymous may identify everyone in it once the sound is on.

Read the edges. Captions burned into the video, usernames in watermarks, a second person visible for a single second, a screen in the background.

Automated tools that examine video generally sample frames at intervals and may not analyse the audio track at all. Treat a "clean" result as meaning that nothing explicit was seen in the frames that were checked, and no more than that.

Ownership

A submitter can only give you permission for something that is theirs. Photographs taken from someone else's social media account, professional photographs, and images of artwork belong to other people. For a small community page the practical risk is usually a complaint and a removal rather than anything worse, but lifting a photo from a private account also raises every privacy problem already discussed. Memes and widely circulated public images are a different matter and are generally tolerated.

常见问题

发布聊天截图前我应该检查什么？

联系人的姓名和照片的标题、每条消息旁边的头像、显示为电话号码而非姓名的信息、消息中使用的姓名、状态栏、任何通知横幅，以及侧面可见的其他聊天。如果对话对故事很重要，更安全的做法是请提交者用自己的话重新讲述。

AI 审核能发现图片中的隐私问题吗？

不能。自动筛选在判断图片是否包含色情或暴力内容方面相当出色，但它根本不检查隐私问题。对于内容安全模型来说，一张陌生人身份证的清晰照片是完全安全的图片。这部分检查需要人工完成。

模糊处理人脸或文件就足够了吗？

只有在模糊程度很高的情况下才足够。对人脸进行轻微模糊处理，认识该人物的人仍然可以辨认出来。对于文件，手机的标记笔通常不够：半透明的荧光笔会使文字透出来。请使用实心不透明块或裁剪掉该区域，并以全尺寸检查结果。

如果提交了未成年人的性爱图片，管理员应该怎么做？

不要保存、截图或发送给任何人，包括其他管理员或作为证据。通过托管平台的举报工具和向警方举报，并遵循关于该文件所给出的指示。

发布视频前应该如何检查？

观看所有内容，打开声音，并注意边缘：烧录字幕、水印中的用户名、一闪而过的第二个人。自动化工具通常会间隔采样帧，并且可能不会分析音频，因此干净的结果仅表示在检查的帧中没有发现任何露骨内容。

A pre-publication checklist

• Does the post need this image at all?
• Every face accounted for: the subject, bystanders, reflections
• No names, handles, numbers or profile pictures in screenshots
• No documents, cards, labels or codes legible
• No address, plate or identifiable location linked to a person
• No one under 18 unless you are confident of consent
• Redactions are solid and survive zooming in
• For video: watched to the end, with sound
• Nothing that could be a fabricated image of a real person

If you are still unsure after going through the list, that is your answer. Publish the text and leave the image out.